Jing Tang

Unifying drug‐target affinity prediction and targeted molecular design within a single interpretable framework remains challenging. Many sequence‐based affinity and design methods rely on global target representations without explicitly modeling binding regions, leading to site‐level ambiguity in both screening and design. By contrast, structure‐based methods require high‐quality structural data and are poorly suited to dynamic targets. In this study, a new model named MolDBG is proposed as a unified site‐aware framework that combines affinity prediction, binding‐site identification, and affinity‐conditioned molecular generation within a single architecture. With binding‐site supervision, MolDBG prioritizes interaction‐critical residues before learning drug‐target representations, reducing false positives from misaligned binding sites and improving interpretability. MolDBG achieves competitive performance across all three tasks while enabling site‐specific affinity prediction and interpretable binding‐site discovery. The framework generalizes to structurally elusive targets, including cryptic pockets and intrinsically disordered proteins. Overall, these results demonstrate that MolDBG is a promising framework for molecular design and screening.

Motivation: De novo drug design requires not only promoting desired target activity, but also avoiding undesired target interactions that compromise selectivity and safety. However, existing generative models largely focus on optimizing positive activity, while overlooking negative activity information that could help suppress off-target effects during molecular design. Results: We present ActivityDiff, a classifier-guided diffusion framework for activity-controlled molecular generation. Unlike conventional approaches that rely primarily on positive activity optimization, ActivityDiff explicitly incorporates both positive and negative guidance through separately trained drug–target classifiers. This design enables the model not only to promote desired target activities, but also to suppress harmful off-target interactions during generation. Experiments show that ActivityDiff effectively supports various drug design tasks, including single- and dual-target generation, fragment-constrained dual-target design, selective generation for improved target specificity, and reduction of off-target effects. These results demonstrate that classifier-guided diffusion with explicit negative guidance provides an effective strategy for jointly optimizing efficacy and safety in molecular design. Availability and implementation: The source code can be obtained from https://github.com/e-yi/ActivityDiff.

Intratumoral cellular heterogeneity limits therapeutic efficacy in cancer patients. Although single-cell transcriptomics offers high-resolution profiling, translating these insights into clinical drug response prediction remains challenging. Recently, transfer learning approaches have attempted to predict patient drug response by leveraging pre-clinical data. However, these approaches operate at the bulk level, often masking the cellular heterogeneity essential for prediction. In this study, we propose scTAPE, a disentangled transfer learning framework to predict patient drug response using tumor single-cell transcriptomics. scTAPE follows a pre-training and fine-tuning paradigm. During the pre-training stage, scTAPE uses a disentangled learning strategy to extract intrinsic pharmacological signals masked by confounding factors from the matched bulk and single-cell expression profiles. Subsequently, a supervised drug response model is trained on labeled cell-line data to fine-tune the aligned common embedding, thereby achieving cross-domain generalization to unseen datasets. Experimental results demonstrate that scTAPE successfully predicts drug response across cell-line datasets and two independent clinical cohorts, outperforming state-of-the-art single-cell-based predictors. Furthermore, by analyzing tumor cell subpopulations, scTAPE not only predicts patient drug response to both single and combination treatments but also identifies potential therapeutic agents targeting drug-resistant subpopulations. The implementation of scTAPE is available via https://github.com/xinliangSun/scTAPE.

Motivation: Identifying synergistic drug combinations holds promise for more effective treatment strategies. Recent deep learning methods such as Transformers and Graph Neural Networks have shown improved predictive performance, but most of them integrate drug and cell line representations without explicitly modelling the causal effects of genes in mediating drug responses. Results: We introduce CADS (Causal Adjustment for Drug Synergy), a deep learning framework that explicitly models the gene–drug causal relationships to improve both prediction accuracy and biological interpretability. CADS integrates multi-omics data with a learnable gene-selection mechanism that performs causal backdoor adjustment, enabling both drug synergy prediction and causal gene discovery. Across multiple benchmark datasets, CADS consistently achieves superior performance compared with state-of-the-art drug synergy prediction models. In addition, downstream analyses on case studies demonstrate that the inferred gene causal scores can recover clinically validated cancer-related genes involved in drug combinations. These results demonstrate that explicitly modelling causal genetic effects can enhance the reliability and interpretability of drug synergy prediction. Availability and implementation: The source code of CADS can be found at https://github.com/HuaiwuZhang/causalDC.

Intratumoral heterogeneity drives variable drug responses in cancer. Single-cell RNA sequencing (scRNA-seq) enables characterization of such heterogeneity and prediction of drug response at single-cell resolution. Accordingly, various computational models have been developed to infer drug response from scRNA-seq data. However, their performance, robustness, and generalizability across different biological contexts remain insufficiently evaluated. To address this gap, we benchmarked representative single-cell drug response prediction models using 26 curated datasets comprising over 760,000 cells across 12 cancer types and 21 therapeutic agents. We constructed balanced and imbalanced scenarios to explore model behavior under different label distributions. To address the lack of ground-truth labels in conventional scRNA-seq datasets, we incorporated two lineage-tracing datasets with experimentally validated drug-response annotations, enabling evaluation in a clinically relevant pre-treatment prediction setting. Our results show that prediction performance was markedly higher in cell lines than in tissue samples and varied substantially across datasets, treatment types, and imbalance levels. Under increasing class imbalance, several methods showed reduced performance, and no single method consistently dominated across all evaluation settings. In lineage-tracing datasets, treatment-associated states were more readily distinguished than lineage-defined pre-sensitive and pre-resistant states, while transcriptomic analyses indicated that pre-existing resistance-related signals were detectable but relatively subtle. For scDEAL, which showed competitive performance across several benchmark settings, label-substitution and ablation analyses further indicated that its performance was jointly influenced by training-label definition and architectural design. Application to an independent PDAC organoid dataset showed that scDEAL predictions were associated with treatment-related transcriptional state changes and biological programs. In summary, our study systematically characterizes the performance and limitations of current single-cell drug response prediction approaches across diverse evaluation settings. These results highlight the influence of dataset characteristics, class imbalance, training design, and the strength of resistance-associated signals on model performance, providing a reference for the evaluation and further development of single-cell drug response prediction methods.

Table S5 shows significant KEGG pathways for the gene expression signatures.

Fig. S3 shows Western blot analysis of GSPT1 protein levels in MOLM-13 and MV4-11 cells after CC-90009 treatment.

Fig. S6 shows HSA synergy landscapes illustrating the interaction effects between various drugs and FLT3 inhibitors in a primary AML patient sample.